Flash flood risk mapping using Analytic Hierarchy Process and machine learning: case of Souk-Ahras City, Northeastern Algeria

As the frequency and severity of floods increase, owing mostly to climate change and anthropogenic activities, identifying flood-prone locations is becoming an increasingly critical task. This study applies a new modeling technique for mapping flash-flood susceptibility in the urban basin of Souk-Ahras, Northeastern Algeria. The study area has been frequently affected by flash floods triggered by torrential rains, steep slopes, and high urbanization rates. The methodology used combines the multi-criteria Analytical Hierarchy Process (AHP) with machine learning, represented by the XGBoost Algorithm. Nine flash-flood conditioning factors were considered, including Land Use Land Cover (LULC), Normalized Difference Built-up Index (NDBI), Rainfall, Topographic Wetness Index (TWI), Slope, Elevation, Curvature, distance to road, and Lithology. The model training procedure used 46 flood spots and 109 no-flood points, which were randomly chosen from sites without a flood history. Model validation, represented by the receiver operating characteristic (ROC) curve, revealed that the AHP-XGBoost model achieved an Area Under Curve (AUC) of 84.5%, compared to 80% and 83% for the standalone AHP and XGBoost models, respectively. This clearly shows an optimal performance for the hybrid model considered.